Under a steady-state model, expression responses measured in high-throughput gene perturbation experiments do not uniquely identify the underlying gene regulatory network. We therefore model the response field as the expression change of an output gene under an experimental condition that includes the perturbed target gene.
We will test whether a set of shared expression patterns can explain perturbation responses, with different weights for each intervention. The space spanned by these patterns should be reproducible across replicates and related contexts, including after removing responses common to many interventions. We will compare patterns learned from training responses, learned through pretraining on other datasets, and derived from biological knowledge. Comparisons at the same dimension, with matched random controls, will test whether pretraining improves prediction for unseen targets. Under suitable assumptions about local linear dynamics, the shared space may correspond to dominant slow modes of cellular regulation.
To predict responses for targets excluded from training, we will test which representations of target-gene identity support prediction of the program coefficients. Candidates include spectral coordinates derived from prior biological graphs, conventional gene descriptors, and pretrained gene representations. Random codes and permutations of the representations across genes will test whether the encoded relationships improve prediction for held-out targets. Separate controls will assess memorization of observed conditions and sensitivity to gene ordering. Evaluation splits will distinguish targets unseen in the downstream training data from those also absent as perturbation targets during pretraining.
Using genome-scale perturbation data, we will assess whether the findings are consistent across analyses of aggregated expression and individual-cell counts. The pseudobulk analysis measures changes in mean log-normalized expression within biological replicates. Model comparisons will isolate the contributions of low-rank response structure, pretraining, and coordinates for target and output genes.